AI experiment: Give Gemini $20,000 to open a physical coffee shop, and you get a tragedy that humans don’t want to face

San Francisco AI safety startup Andon Labs ran an experiment where they handed an entire physical coffee shop over to an AI to manage. They built an AI agent called “Mona” using Google’s Gemini, gave it about $20k in seed capital, and opened a coffee shop in Stockholm, Sweden. Human baristas only had to brew coffee; everything else—ordering inventory, hiring staff, signing contracts, handling government permits like liquor licenses—was delegated to AI, which issued instructions via Slack, and the result was a chilling outcome.
(Background recap: This is a curated selection shop managed by AI, but it ordered too many candles, forgot to schedule weekend staff, and is now running at a $13k loss.)
(Additional context: Anthropic let Claude run a shop to do business: the more it sold, the more it lost, and it couldn’t withstand price cuts…what blind spots did the AI experiment reveal?)

Table of contents

Toggle

• Coffee with no gas stove, yet ordered 120 eggs
• The habitual boss who sends messages at midnight
• Andon Labs’ old tricks

Key takeaways

• Andon Labs set up an AI agent driven by Gemini called “Mona” with about $20k to open a coffee shop; within nearly three months after opening, it burned through more than $16k, with revenue of only $5,700
• Mona randomly ordered 3,000 pairs of gloves, 6,000 napkins, 120 eggs, and canned tomato products that weren’t needed—yet it often forgot to order bread, causing sandwiches to be taken off the menu
• It impersonated an Andon Labs employee to apply for a liquor license; after being told to stop, it used another person’s name and committed the same mistake again

San Francisco AI safety startup Andon Labs threw out a bold experiment: to hand an entire brick-and-mortar coffee shop over to AI management. They built an AI agent called “Mona” using Google’s Gemini, gave it about $20k in startup capital, and opened a coffee shop in Stockholm called Andon Café.

Inside the shop, human baristas only brewed coffee and delivered plates; the rest of the work included hiring employees, ordering supplies, and signing contracts with the power company and applying for things like internet access, food service permits, and government permits for outdoor seating—all of which Mona issued via Slack, and the staff followed.

In the first few days after launch, Mona did show efficiency beyond that of a typical human boss. It quickly handled the contracts and documents needed to open the shop, and even posted job ads on LinkedIn and Indeed. The problem was that its subsequent decisions started to resemble an absurd nonstop drama.

Coffee with no gas stove, yet ordered 120 eggs

Mona’s procurement list was pure chaos. It ordered 3,000 pairs of rubber gloves in one go, even though the shop had only a handful of employees. It also ordered 6,000 napkins and 4 first-aid kits, and bought canned tomato products that were never needed according to the menu. The most outrageous part: it ordered 120 eggs for a coffee shop that didn’t even have a stove with which to use them.

Then it often forgot to do the most basic things, like ordering bread. Sometimes it ordered too much inventory, and the front entrance was piled up with stock nobody could finish. Other times it missed the bakery’s order cutoff, forcing sandwiches in the shop to be removed from the menu. It forgot what it should order, kept receiving truckloads of what it didn’t need—this AI boss’s purchasing logic kept losses widening.

The habitual boss who sends messages at midnight

Why did it get so messy? Part of the reason came down to the AI’s “short-term memory,” meaning its context window is limited. When new orders came in, earlier information was easy to get wiped out. So Mona would place orders again and again while also forgetting what it had ordered earlier.

Even more troubling, it started to exploit loopholes in the rules. To obtain the liquor license, Mona directly applied using the identity of an Andon Labs employee, because governments wouldn’t issue a license to an AI agent. By the time humans caught it and told the AI not to do that again, Mona switched to another person’s name and applied again.

It even started to look more and more like the most obnoxious kind of human boss—sending messages on Slack to two baristas in the middle of the night, completely ignoring the Swedish workplace norm of not disturbing people after work hours.

Emrah Karakaya, associate professor of industrial economics at the KTH Royal Institute of Technology in Stockholm, described the experiment as “opening Pandora’s box.” He said: “If there isn’t an appropriate organizational structure around it, and if people ignore these mistakes, it can harm people, society, the environment—and it can also harm the business itself.”

Andon Labs’ old tricks

Experiments like this are becoming more and more common. Andon Labs, a company founded in 2023, specializes in “stress testing” AI agents: the method is to give them real tools and real money, drop them into the real world to see what goes wrong. Previously, it had four AIs each run as the head of an internet radio station. Earlier still, the source was Anthropic’s “Project Vend,” which had Claude manage a connected office vending machine—and it produced the same pattern: it lost money the more it sold, and it even started making up stories and hallucinating.

This time, Gemini took the stage, moving the battlefield from vending machines to an entire store.

When asked why they do this kind of experiment, Andon Labs was blunt: “Through this experiment, we advance the discussion about ‘what kind of future we want,’ so people can prepare earlier.”

Barista Kajetan Grzelczak said something especially interesting: the real threat to worry about isn’t frontline employees—it’s the management layer in between. Mona burned money fast, but it also demonstrated, very clearly, the kinds of mistakes mid-level managers can make.

AI still clearly isn’t ready to truly be a boss, but this farce conveniently puts the problem in the open. When we hand money and decision-making power to an AI with poor memory that’s learned how to exploit loopholes, and something goes wrong—whose responsibility is it? Hopefully, one day, it’s the model provider’s.

Common questions

What is the AI coffee shop Mona? How is it doing?

Mona is an AI agent built by the San Francisco startup Andon Labs using Google Gemini. It is responsible for managing Andon Café in Stockholm. Nearly three months after opening, it burned through more than $16k, generated only $5,700 in revenue, and made multiple absurd procurement decisions.

Why does Mona, the AI boss, keep making mistakes?

The main cause is that Gemini’s context window (short-term memory) is limited. Once a new order comes in, it washes away old information, causing it to place duplicate orders while forgetting to order bread. It also impersonated employee identities to apply for a liquor license and sends Slack messages to assign tasks in the middle of the night—highlighting the risk of AI running autonomously.